You are on page 1of 24

Journal of Phonetics (2002) 30, 139–162

Available online at on

The phonetics of phonological speech errors:

An acoustic analysis of slips of the tongue
Stefan A. Frisch
Department of Communication Sciences and Disorders, University of South Florida,
4202 E Fowler Ave, PCD1017, Tampa, FL 33620, U.S.A.

Richard Wright
Department of Linguistics, University of Washington, Box 354340, Seattle,
WA 98195-4340, U.S.A.

Received 22nd November 2000 and accepted 7th February 2002

Acoustic analysis was used to examine whether speech errors involve

lexical, segmental, or sub-featural errors in speech production. Nine
participants produced tongue twisters that induced errors between /s/
and /z/ word onsets in contexts where the error outcomes were either
words (e.g., sit to zit) or nonwords (e.g., suck to *zuck). Three
measurements of the /s/-/z/ contrast were made: (1) percent voicing,
(2) duration of frication, and (3) amplitude of frication. The tokens
were also transcribed under careful listening conditions. Gradient and
categorical errors were found for all acoustic dimensions. The errors
might or might not be detected by careful listening, depending on the
extent to which there were errors along all three dimensions. These
data support previous articulatory studies that found speech errors at
a sub-featural level. However, cases where /s/ and /z/ are realized
with a categorical change in voicing are more common than would be
expected if categorical changes in voicing were merely extreme
examples of gradient voicing errors. Also, both gradient and
categorical error rates were higher when the error outcomes were
words. Thus, our study also provides evidence for the psychological
reality of phonological segments and words as units in the speech
production process. r 2002 Elsevier Science Ltd. All rights reserved.

1. Introduction

Speech errors have traditionally been used to provide evidence for models of speech
production that utilize the constructs of linguistic theory as psychologically real
components of linguistic performance (e.g., Levelt, 1989). While it is indisputable
that speech errors do occur, few unambiguous conclusions about the mechanisms

Address correspondence to: S. A. Frisch. E-mail:

0095–4470/02/$ - see front matter r 2002 Elsevier Science Ltd. All rights reserved.
140 S. A. Frisch and R. Wright

of speech production can be drawn from speech error data. In addition, several
researchers have questioned the validity of phonological speech error data that has
been recorded using phonetic transcription (Laver, 1980; Mowrey & MacKay, 1990;
Boucher, 1994; Ferber, 1995). In this paper, we undertake an acoustic analysis of
speech produced by nine talkers in a speech error elicitation experiment. Acoustic
analysis circumvents the problems of perceptual bias introduced by phonetic
transcription. Our analysis provides evidence for the psychological reality of
phonological segments in speech production as a statistical tendency, supporting
transcriptional analyses. However, we also find evidence for speech errors at a sub-
featural or gradient phonetic level that have not previously been attested. Our data
support a model of speech production where individual gestures are organized into
gestural constellations at the level of the segment (Saltzman & Munhall, 1989; Byrd,
1996). Segments are further organized into words. We find that the segment and
word levels influence the implementation of gestures in both erroneous and error-
free productions.

1.1. Background
Phonological speech errors (also called sub-lexical errors) have been an important
source of evidence for the psychological reality of phonological features and
segments. In many speech errors, it appears that portions of the intended utterance
are produced in an unintended order. It is claimed in the speech error literature that
the misordered portions correspond to abstract linguistic units such as onsets, codas,
phonemes, segments, and features. The errors in (1) are given by Fromkin (1971) as
support for the psychological reality of segments, distinct from words or syllable

(1) a. frish gotto ‘‘fish grotto’’

b. blake fruid ‘‘brake fluid’’
c. spicky point ‘‘sticky point’’

In (1a), the /r/ in grotto is presumably misordered, and appears as part of

the preceding word instead. In (1b) the /l/ and /r/ are exchanged, each appearing in
the others’ place. In (1c), the /p/ (or alternatively, the [+labial] feature) is
anticipated, but also repeated in its proper place. As with (1c), it is often the case
that any particular error can be interpreted in more than one way. Another error
from Fromkin (1971), glear plue sky for clear blue sky, is claimed to involve an
exchange of the voicing feature. This demonstrates that errors involving linguistic
features are possible and thus that features are psychologically real units of
processing as well.
Mowrey & MacKay (1990), using electromyographic (EMG) recordings of tongue
twister production conclude ‘‘that errors which have been consigned to the
phonemic, segmental, or feature levels could be reinterpreted as errors at the motor
output level’’ (p. 1311). In the remainder of this section, we review the
transcriptional methods of speech error analysis and the results of the instrumental
study of Mowrey & MacKay (1990).
The phonetics of phonological speech errors 141

1.2. Transcriptional approach

Traditional approaches to speech error analysis use phonetic transcription to encode

speech errors at the time they are heard. In ‘‘naturally occurring’’ speech error
corpora, errors that are observed in everyday speech are written down opportunis-
tically. In the early corpora (e.g., Fromkin, 1971; Shattuck, 1975) the error recorders
were usually participants in the communicative event in which the error occurred.
Stemberger (1983) collected naturally occurring errors only as an observer in an
attempt to reduce the potential for perceptual bias. In some cases, recordings of
naturally occurring speech are used, and suspected errors are listened to repeatedly
to ensure accurate transcription (e.g., Garnham, Shillcock, Brown, Mill & Cutler,
1982). Transcription is also normally used to encode errors in speech error
elicitation experiments (e.g., Baars, Motley & MacKay 1975; Dell & Reich, 1980),
though usually the utterances themselves are recorded on tape or computer and
listened to repeatedly.
In all cases where transcription is used, the noting of a speech error necessarily
coincides with the hearer noticing an anomalous percept. Thus, in transcriptional
approaches, a speech error is defined to be an utterance that produces an anomalous
percept that would be recognized as anomalous by the speaker (Dell, 1986). Mowrey
& MacKay (1990, p. 1299) note that imperceptible speech errors may also exist and
claim that ‘‘such production anomalies are errors if speech output differs from the
speaker’s intended output, however subtle the anomaly’’. Their claim raises the
question of how articulatorily detailed the speaker’s intentions are, which we discuss
Transcribed speech error evidence has been used to argue in favor of the
psychological reality of many phonological units, including the feature, segment,
phoneme, cluster, syllable, and word. Among sub-lexical errors it has been claimed
that errors occur primarily at the level of the phoneme or feature (Wickelgren, 1965)
and that erroneous utterances are phonetically and phonotactically grammatical
(Wells, 1951; Fromkin, 1971). In other words, it is claimed that speech errors occur
by misordering abstract phonological units and the result is a phonetically normal
segment and possible word according to the grammar of the language. Phonetic
errors are often explicitly argued against (e.g., Fromkin, 1971) and it is claimed that
when abstract units move to different locations, they phonetically accommodate to
their new environment. It should be noted that there is some disagreement on these
conclusions among experimenters using the same collection techniques. For example,
Stemberger (1983), based on his own corpus of naturally occurring errors, claimed
that phonologically ungrammatical utterances do occur, though infrequently.
The use of transcription to encode speech errors has received widespread criticism
and been the subject of some empirical research (Laver, 1980; Garnham et al., 1982;
Shattuck-Hufnagel, 1983; Mowrey & MacKay, 1990; Boucher, 1994; Ferber, 1995).
There are two primary criticisms. First, the use of phonetic transcription cannot
capture sub-contrastive or gradient errors, below the level of a segment or feature,
since the transcription system is inherently segmental. If gradient errors do occur,
careful transcription of repeatedly heard recordings of speech errors would probably
discover some of them. However, speech errors heard in conversation are usually
only broadly transcribed and the listener’s full attention is not on phonetic detail.
Thus the errors contained in naturally occurring corpora (Fromkin, 1971; Shattuck,
142 S. A. Frisch and R. Wright

1975) may only represent a portion of the actual speech errors produced in natural
dialogue, and any model based on transcription evidence is therefore unable to
answer questions about the phonetic details of speech errors. Errors collected in
speech error inducing experiments (e.g., Baars et al., 1975; Dell, 1986; Shattuck-
Hufnagel, 1992) might be more revealing of phonetic detail, since the errors are
recorded and can be reviewed many times over. However, the design of these
experiments is usually to produce a specific error. Thus, the experimenter’s
transcription task is a forced-choice decisionFis it an error or notFrather than
an unconstrained phonetic transcription task.
The second criticism of error collection using transcription is that the transcript is
subject to the perceptual biases of the listener. It is well known from the literature
on speech perception that speech is perceived in the context of the language system
of the listener (see Wright, Frisch & Pisoni, 1999, for a recent review). For example,
in the phenomenon of categorical perception, phonetically anomalous speech sounds
that are acoustically intermediate between two categories are perceived by naive
listeners as members of one category or the other, rather than a blend (see
Liberman, 1997, for several articles). In another phenomenon, known as phonemic
restoration, speech samples that have had segments replaced by noise or a cough are
perceived as intact. Listeners, even when informed that there is a missing segment,
are unable to accurately report which segment is missing or where in the word the
disruption occurred (Warren, 1970; Samuel, 1981). Research on the detectability of
mispronunciations of segments in running speech has found that the likelihood of
detecting an error depends on the error’s place within the word and sentence, and
the predictability of the word in its sentential context (Cole, 1973; Marslen-Wilson
& Welsh, 1978). In summary, speech error percepts are subject to systematic biases
and it is unclear whether many of the patterns observed in transcriptional speech
error analyses are informative of linguistic biases in the speech production process
or merely a reflection of the hearer’s perceptual system.

1.3. Instrumental data

Mowrey & MacKay (1990) present an electromyographic (EMG) study of sub-
lexical speech errors elicited using tongue twisters. They used EMG recordings of
the orbicularis oris muscle (lower lip) and lingual transversus/verticalis muscle
(tongue blade) in combination with audio recordings to determine whether
noncontrastive errors occur. They acted as their own subjects. The tongue twisters
they used, shown in (2), crucially involved segments with lower lip (/P/ in 2a) or
tongue blade (/l/ in 2b, c) articulations that were in proximity to segments that did
not contain such articulations (/s/ and /r/ or nothing, respectively).

(2) a. She sells sea shells by the seashore

b. Bob flew by Bligh bay
c. Fresh fried flesh of fowl

Mowrey and MacKay interpreted unexpected muscular activity as evidence for an

error in speech production, and so in their study the definition of an error is
crucially different from a perceptual definition based on transcription. They found a
The phonetics of phonological speech errors 143

number of instances of inappropriate muscular activity where there was no

perceptible anomaly. In addition, they found several cases of inappropriate muscular
activity where there was a potentially intermediate percept that was not a clear
member of either category. They conclude based on this evidence that gradient
errors do occur, and they further claim that such errors are quite frequent. In one
set of 150 recordings of Bob flew by Bligh bay they report 48 tokens containing
errors involving intrusive transversus/verticalis muscle activation, the majority of
which involved an amount of activation intermediate between none and that which
is appropriate for a normal [l] production. Their proposal, from an articulatory
perspective, is that sub-lexical errors occur on a continuum of gestural activity, and
are neither segmental nor grammatical under any reasonable definition of these
terms. Preliminary data from a recent study of one speaker using magnetometry
supports the claim that gradient activation results in partial or incomplete gestures
(Pouplier, Chen, Goldstein & Byrd, 1999).
These articulatory studies of gestural activity during tongue twister production
clearly reveal that there can be gradient levels of gestural activation that have
imperceptible consequences. Mowrey and MacKay conclude that such evidence
demonstrates that

neither the speaker nor the listener may be aware of the true nature of the errors made or
whether indeed an error has been made at all y [which] contradicts any model claiming
that ‘‘low-level’’ phonetic processes are necessarily overseen by a higher-order segmental or
featural planning unit (p. 1311).

While transcription is too coarse a means of encoding the phonetic details of an

utterance, these instrumental studies have yet to explore how these gradient errors fit
into the broader pattern of speech production. It is possible that other muscles were
also gradiently activated in a coordinated fashion during these errors. Patterns of
muscular coordination in speech errors would provide evidence of organization at
the level of the feature or segment in speech production. If physically independent
articulations for a segment are found to coordinate in gradient errors, such as lip
rounding and tongue blade raising for /P/, then the most plausible explanation
would be the partial activation of a higher-order segmental planning unit. It is also
possible that the behavior of other muscles in speech errors is independent. The
available articulatory evidence to date demonstrates that gradient speech errors can
and do frequently occur at the sub-segmental level, but this data does not rule out a
role for higher levels of organization in speech production as a constraint on speech
It is also worth noting that defining an error as unexpected muscle activity
provides no allowance for what might be normal variation, by assuming that the
speaker’s intention excludes such activity. These articulatory studies have found
many of these gradient errors to be imperceptible, even upon repeated listening with
high-quality equipment. It may be that these gestures have no significant acoustic
effect. A possibility that should be considered is that one goal when talking is to
convey a distinct, contrastive message to an observer. Thus, random variation that is
noncontrastive may not be monitored as strictly and for this reason such produc-
tions might not be considered erroneous by the producer or detected as erroneous
by the monitoring mechanisms in the production system of the producer.
144 S. A. Frisch and R. Wright

1.4. Overview of the study: acoustic analysis of speech errors

Phonetic transcription provides a coarse coding of speech, filtered through the
perceptual system of the listener. EMG analysis of a single muscle provides detailed
information about activity levels of an individual articulator but little information
about the coordination of articulators or the acoustic effects of the articulation. We
propose to analyze experimentally elicited speech errors using acoustic measures.
Acoustic analysis can examine both contrastive and sub-contrastive dimensions
together, providing an intermediate level of analysis between transcription and
EMG. In addition, acoustic analysis may reveal systematic biases in the perception
of speech errors. Acoustic analysis has the further benefit that larger numbers of
experimental subjects can be studied. The following section contains a description of
the data and analysis techniques used in our acoustic analysis of phonological
speech errors.

2. Methods

2.1. Data
The data for our analysis come from a recorded corpus of utterances from a speech
error experiment that induced errors using tongue twisters (Frisch, 1996, Chapter 9).
In the experiment, 21 participants each produced 6 repetitions of 88 different tongue
twisters. The participants were monolingual American English speaking under-
graduate students at Northwestern University who were paid for their participation.
The experiment was self-paced and the speech rate of the participants was not
controlled. The tongue twisters were printed individually in large type on index
cards. The participants read each tongue twister aloud three times, and then
repeated the tongue twister from memory three times. If they forgot the correct
words, participants were allowed to consult the index card between repetitions
during the repeat-from-memory portion. Each participant had a break after half of
the experiment was completed in which they were engaged in normal conversation
for about 5 min. The experiment lasted about 45 min in total. The entire session was
recorded on audiocassette using a Marantz portable cassette recorder. The
participants wore an electret condenser microphone attached to their shirt in the
upper chest area.
The original experiment was a psycholinguistic study designed to induce word
onset consonant errors for 22 different consonant pairs. Each consonant pair was
used in four different tongue twisters. Each tongue twister consisted of four
monosyllabic words. We examined the /s/-/z/ tongue twisters from the experiment,
given in (3). Two of the tongue twisters created target errors that were existing
words (3a, c), and two created nonword target errors (3b, d).

(3) a. sit zap zoo sip

b. sung zone Zeus seem
c. zit sap sue zip
d. zig suck sank zilch
The phonetics of phonological speech errors 145

The words in the tongue twister and their error targets (if they were words) were
balanced for lexical frequency within each tongue twister. Thus, lexical frequency
effects should not differentially influence the relative error rate between the
consonants, from intended /s/ to [z] or intended /z/ to [s] in this particular case.1
The acoustic analysis focuses on the single consonant pair, /s/-/z/. Data from nine
participants in the original experiment were analyzed. They were the first nine that
participated in the experiment. Twisters containing /s/-/z/ were selected for this study
as the first author found some of the tokens in these twisters to be perceptually
ambiguous when scoring the original experiment. It was assumed that these stimuli
had good potential for finding a variety of error types and examining the
distribution of both categorical and gradient speech errors. The choice of the first
nine participants in the experiment was arbitrary. Five of these participants were
male and four were female. These participants produced a total of 448 /s/ tokens
and 446 /z/ tokens. The number of productions of each segment in each word by
each participant is not necessarily equal, as the participants sometimes repeated
individual words or the entire tongue twister after producing an error. All
productions were measured, including repetitions.
It should be noted that the use of nonsense tongue twisters to generate speech
errors may introduce systematic patterns not observed in naturally occurring speech,
and thus the results of this paper may not generalize to normal speech production.
Shattuck-Hufnagel (1992) compared perceived error rates in nonsense tongue
twisters with those in spontaneously generated utterances using the same words.
There was a difference in the number of errors observed in the two conditions, but
the relative patterns of errors were the same for both conditions. Since quantitative
but not qualitative differences were found, it appears that speech in tongue twisters
does not differ in important ways from spontaneously generated utterances. Note
that Shattuck-Hufnagel’s (1992) study used transcribed speech error evidence. It may
be that tongue twisters enhance the likelihood of observing gradient speech errors,
given that they involve an unusually high level of repetition and alternation.
However, we assume that this entails quantitative rather than qualitative differences
in the action of the production mechanism, and that the same phonological
encoding and phonetic implementation is involved in producing both tongue twisters
and natural speech.

2.2. Measurements
The onset /s/ or /z/ of each word was measured in a variety of ways, with the
overall goal of finding quantitative differences between /s/ and /z/ along acoustic
dimensions that can be plausibly associated with independent aspects of the
articulation. In describing the data acoustically, tokens are categorized as intended
/s/ or /z/ based on the word the speaker was supposed to say if the tongue twister
was to be produced correctly. This was assumed to be the speaker’s intended
utterance. Both perceptually normal and erroneous tokens are included in the figures
and tables that follow.

For convenience, we borrow the slash and bracket notation of phonology to differentiate the intended
production from the recorded utterance. The ‘‘underlying’’ /s/ and /z/ represent the intended productions
of the participants, and the ‘‘surface’’ [s] and [z] (or other transcriptions in square brackets) represent the
actual output.
146 S. A. Frisch and R. Wright

Three measures are reported in this paper. These measures differentiate /s/ and /z/
tokens produced by our talkers and have articulatory bases that are sufficiently
physiologically independent for our analysis. In addition to measuring each token,
we transcribed each token based on repeated careful listening to digitized recordings
of the individual repetitions of the tongue twisters. In most cases, the result was a
clear /s/ or /z/ percept. However, there were several tokens of /z/ that were trans-
cribed as devoiced.
The twister sung zone Zeus seem contained a sequence of coda /s/ followed by
onset /s/. In 40 of the 55 repetitions of this twister, the participants did not produce
two distinct /s/ segments (though in most cases there were two distinct amplitude
peaks). These ‘‘geminate’’ /s/ tokens were not included in the analysis as they could
not be unambiguously measured. In addition, there were 10 instances of ‘‘concat-
enation’’ errors where the participant produced a long [sz] or [zs] sequence,
three instances where [W] was produced in place of [z], and three instances that
were perceptually bizarre and untranscribable. We discuss these tokens in
Section 3.3, and they are excluded from our analysis in order to provide the
clearest possible picture of normal /s/ and /z/ production by our participants. Given
these criteria, data for 397 /s/ productions and 435 /z/ productions are reported in
this section.
The first measure is the duration of the fricative. We considered the duration of
the fricative to be the total duration of fricative noise, including any measurable
overlap of the frication noise with the formant structure of the following vowel and
the preceding vowel or sonorant, if any. In general, this overlap was common for /z/
but lacking for /s/. We determined whether or not fricative noise was present by
examining the waveform for the fuzziness of aperiodic noise and a spectrogram for
evidence of broadband noise at high frequencies. The expected pattern was for /s/ to
be longer than /z/ (Klatt, 1976), which was true across all nine talkers in our study.
Fig. 1 shows means and error bars representing one standard deviation for each
participant for our three measures of /s/ and /z/. Average duration is given in the
top panel of Fig. 1, with the average for /s/ shown by the shaded bar, and the
average for /z/ shown by the clear bar.
The second measure is the amplitude of frication noise. Amplitude of frication
noise was calculated based on the RMS amplitude of the signal after high-
pass filtering at 2 kHz. For each fricative, RMS amplitude was computed for
10 ms windows across the entire fricative. The window containing the peak
RMS amplitude was identified, and the amplitude of frication was defined to
be the RMS amplitude of the fricative in a 50 ms window around the peak.
This generally corresponded to the middle of the fricative segment. The expected
pattern was for /s/ to have more frication noise than /z/ (Strevens, 1960; Pickett,
1980), which was true across all nine participants in our study (see Fig. 1, middle
The final measure is the percent of the fricative that contained voicing. We used
the waveform to determine when voicing was present, and considered portions of
the signal that had evidence of clear periodicity as voiced. The percent of voicing
was the fraction of the total duration that contained voicing. In most cases, there
was clear evidence of a glottal closure phase that indicated the presence of voicing.
In some cases there was very breathy voicing resulting in sinusoidal waveform
overlaid with frication noise. Breathy voicing was considered voicing until the point
The phonetics of phonological speech errors 147

where the pressure valleys marking each glottal cycle were too obscured by frication
noise to be reliably identified. The expected pattern was for /z/ to have a larger
percent of voicing than /s/, which was true across all nine talkers in our study (see
Fig. 1, bottom panel).
The measurement of duration, frication amplitude, and percent voicing for a
sample case of /z/ is shown in Fig. 2. This is a production of zone where fricative
noise overlaps with the coda nasal of the previous word (sung). The fricative is
devoiced in the middle, but voicing begins again before the vowel onset. This
pattern is relatively common for /z/ (Haggard, 1978; Smith, 1997). The top panel in
Fig. 2 shows the original signal, with brackets indicating the overall length of
frication and the portions that are voiced. The bottom panel shows the signal
filtered at 2 kHz, which has removed most of the influence of periodicity, and the
50 ms window around the peak over which the RMS amplitude of frication was

Figure 1. Mean duration (top panel), frication amplitude (middle panel),

and percent voicing (bottom panel) for each participant’s productions of
/s/ and /z/. Error bars indicate one standard deviation around the mean.
148 S. A. Frisch and R. Wright

Figure 2. Example measurement of /z/ in zone showing waveform and high-

pass filtered waveform. Relevant regions for measuring duration, frication
amplitude, and percent voicing are indicated by braces.

While our goal was to consider measurements that are articulatorily indepen-
dent, it is clear that duration, amplitude of frication, and percent voicing are not
unrelated aerodynamically (Ohala, 1983). During voicing, amplitude of frication is
necessarily reduced, as the closed phases of the glottal cycle stop the airflow that is
crucial for maintaining frication. However, many speakers produce /z/ with a voice-
less portion in the middle of the fricative. During this devoiced portion, airflow is
not constrained by glottal closure. Our measure of frication amplitude over the
loudest 50 ms window would allow the frication amplitude of /z/ to be determined
by the devoiced portion, if there is one. In addition, talkers also actively increase the
pulmonic airflow in /s/ to produce a noisier fricative (Shadle, 1985). In an error
articulation, the oral and glottal gestures for /s/ may be in place, but if this is not
combined with increased pulmonic airflow, the resulting frication amplitude will be
less than normal for /s/.
Duration may also interact with voicing. During a fricative, pressure buildup
behind the oral constriction equalizes transglottal pressure, making voicing more
difficult. Thus, long duration voiced fricatives may have a stronger tendency to
devoice. However, the connection is not an absolute one, as it is perfectly possible
to maintain voicing during a long fricative. Though rare, voiced geminate fricatives
are attested in the world’s languages (e.g., Italian). Note also that this dependence
only predicts that long /z/ will devoice, not that a very long fricative is necessarily
entirely voiceless. We feel that the dependence between acoustic properties measured
in our study cannot wholly be explained by aerodynamic interdependence between
our measures, and variation is to some extent under the active control of the talker.
Thus, there is room for more abstract levels of linguistic organization to influence
the interrelation of our measures.
The phonetics of phonological speech errors 149

3. Categorical, gradient, and ungrammatical errors

In the previous section, it was demonstrated that the three dimensions of duration,
frication amplitude, and percent voicing differentiate /s/ and /z/ for our participants.
Of these, the percent voicing measure most directly reflects the voicing contrast.
There is considerable overlap between /s/ and /z/ in the distribution of frication
amplitude and duration, even among tokens that were perceived to be correctly
produced. As a result, the effects of frication amplitude and duration are likely to be
secondary in influencing the percept, and less definitive in determining when an error
has occurred (Baum & Blumstein, 1987). In this section, we first investigate variation
in percent voicing, and then consider the variation in frication amplitude and
duration. Finally, we briefly discuss the other disfluent productions found in our
corpus. Overall, the presence of these strange utterances and the considerable
variability in production even for percent voicing supports the claim that many
ungrammatical speech error outputs occur (Mowrey & MacKay, 1990).

3.1. Variation in percent voicing

The percent voicing of all tokens for each participant is shown in Fig. 3. The data
are grouped to illustrate productions that are usual and unusual for each
participant. Both 0 and 100% voicing are given separately. The other groups are
0–5, 5–10, 10–30, 30–60, and 60–100%. Each panel of Fig. 3 shows the number of
occurrences for one participant. Within each panel, the shaded bars indicate
occurrences for intended /s/ and the clear bars indicate occurrences for intended /z/.
The data in Fig. 3 are summarized in Table I, where the aggregate number of
occurrences for each percent voicing category for all participants is given for both
/s/ and /z/.
Given that these tongue twisters were recorded in an experimental laboratory
setting, we assume the participants were hyperarticulating and attempting to produce
clear speech. Thus the the target percent voicing is likely to be 0% for /s/ and 100%
for /z/. In normal speech, partial devoicing of /z/ results in productions that are
voiceless in the middle, though these productions are usually still more than 60%
voiced (Haggard, 1978; Smith, 1997). Fig. 3 and Table I show that many
productions did not fall into the percent voicing groups that would normally be
predicted for /s/ and /z/. Utterances where the percent voicing was between 5 and
30% are certainly anomalous for both /s/ and /z/. These productions with inter-
mediate amounts of voicing provide evidence that gradient, noncategorical errors
are made (Mowrey & MacKay, 1990).
Overall there are fewer cases of intermediate voicing for /s/ than for /z/.2 Among
the utterances with intermediate voicing, /s/ also shows a clear pattern of variability.
As the percent voicing increases, fewer and fewer tokens of /s/ are found that have
that degree of voicing. This is the pattern we might expect for phonetic variability if

Participant 6 is a curious exception to the generalizations about the likelihood of gradient errors in /s/
and /z/. For participant 6, /s/ is more variable, and was overlapped with the following vowel (and the
preceding vowel if there was one) a small amount in many utterances. On the other hand, /z/ for
participant 6 is much less variable, with almost all instances realized with 100% voicing. Participant 6 is
also atypical in the production of categorical errors, and does not follow the /s/ or /z/ pattern of the
majority of the participants as discussed below.
150 S. A. Frisch and R. Wright

Figure 3. Distribution of tokens among percent voicing groups for /s/ and /z/
for each participant.

TABLE I. Distribution of intended /s/ and

/z/ tokens by percent voicing for all

Intended segment

Percent voicing /s/ /z/

0% 324 56
0–5% 25 8
5–10% 13 24
10–30% 12 39
30–60% 4 23
60–100% 6 33
100% 13 252
Total 397 435
The phonetics of phonological speech errors 151

0% voicing is intended. Productions that are farther away from the intended are
relatively less likely, as they lie on the tail of the distribution. By contrast, the
devoicing of /z/ was relatively common and does not clearly reflect degrees of
variation away from an intended production of 100% voicing. Levels of voicing
between 5 and 100% occur with relatively equal frequency and there are no
apparent systematic patterns. It may be that, since the devoicing of /z/ is common in
normal productions, the data here are mixture of normally devoiced /z/ and cases
where /z/ is partially devoiced in error.
Fig. 3 and Table I also provide distributional evidence for categorical errors.
Across all participants, there are 18 cases where /s/ was produced with more than
60% voicing (out of 397 total). Compared to the five cases with 30–60% voicing
and 12 cases with 10–30% voicing, the number of completely voiced tokens is
relatively large. For the individual data of six of nine participants the number of
tokens with more than 60% voicing was greater than the number of tokens with
10–60% voicing. This suggests that in addition to the peak in the distribution of /s/
at 0% voicing, there is a second peak in the distribution of /s/ at 100% voicing. We
interpret this second peak to be the result of categorical changes from intended /s/
to [z].
There is also phonetic evidence that the tokens of /s/ with greater than 60%
voicing are categorical errors. These tokens have shorter duration and lower
frication amplitude than the typical /s/. In addition, these tokens have overlap
between the frication noise and the onset of the formant structure of the follo-
wing vowel that is typical of /z/ in our data. Finally, these fricatives have ampli-
tude envelopes that gradually increase from the amplitude of the fricative toward
the amplitude peak of the following vowel, which is also typical for /z/ in our
There are 56 cases where /z/ was produced with 0% voicing (out of 435 total).
The number of completely voiceless cases is larger than the number of cases with
low levels of intermediate voicing (0–10%) for the individual data of seven of the
nine participants. In about half of these cases, these also appear upon inspection to
be categorical errors, with /z/ produced as [s]. In these cases, there is /s/-like
duration and frication amplitude. There is a small positive VOT between the offset
of /z/ and the onset of voicing and formant structure for the following vowel. There
is also an abrupt decrease in frication amplitude of the /z/ followed by an abrupt
rise in vowel amplitude to the vowel amplitude peak that is characteristic of /s/ in
our data. However, there are other cases of 0% voicing that lack these additional
/s/-like features where the result may be better transcribed as ½z .
One example of an apparent categorical intended /z/ to [s] error is given in Fig. 4.
The top panel of Fig. 4 shows [s] in the production of zap by participant 5. The
bottom panel shows an error-free production of sap by the same participant. In
both utterances, there is a relatively large amount of frication noise, and the offset
of the fricative is complete before the onset of the following vowel begins.
An example of an error where intended /z/ is realized as ½z  is given in Fig. 5. The
top panel of Fig. 5 shows ½z  in the production of zap by participant 9. The bottom
panel of the figure shows an error-free production of sap by the same participant. In
contrast with the categorical example above, the ½z  has a much lower frication
amplitude, and the frication noise offset is at about the same time as the vowel
onset. So while there is no measurable voicing during the fricative, there is no delay
152 S. A. Frisch and R. Wright

Figure 4. Example categorical error in the production of zap (top panel) and
normal production of sap (bottom panel) for participant 5.

Figure 5. Example devoicing error in the production of zap (top panel) and
normal production of zap (bottom panel) for participant 9.

in voice onset as there is with /s/. Thus, it appears that even though there is no
voicing, this error is not a categorical change from /z/ to [s].
The variation in percent voicing suggests that more errors (whether gradient or
categorical) were from /z/ toward [s] than from /s/ toward [z]. This pattern is
expected, as /s/ is higher in frequency of occurrence than /z/ and in general low-
frequency items are replaced by high-frequency items in speech errors (Stemberger,
1983). However, it has been reported in the speech error literature that /s/ and /z/
errors are one of several cases where the normal frequency effect is reversed
(Stemberger, 1991). Stemberger found that /s/ to [z] errors were more common than
/z/ to [s] errors in a transcribed speech error experiment designed to examine asym-
metries in segmental errors.
The phonetics of phonological speech errors 153

Though the phonetic variability suggests that /z/ errors are more common than /s/
errors, variability in voicing in /z/ does not affect the perception of /z/ nearly as
much as it affects the perception of /s/. Fig. 6 shows our judgement of the error rate
based on our percept of each token. Tokens are aggregated into the same percent
voicing groups as above, and the percent that were perceived as errors is plotted
against the average percent voicing of tokens in the group. Our judgement of the
percept was determined informally by consensus, based on repeated listening
through headphones at a computer workstation. We found our level of agreement
on the percept to be high. A token was considered to be an error if it was an /s/
that was not perceived as [s] or if it was a /z/ that was not perceived as [z]. While
we did perceive some cases as ½z , many cases that would be best described as ½z 
were not perceived to be anomalous. These cases were only identified as ½z  by visual
inspection of the waveform. In the perceptual data, only the cases of perceived ½z 
were counted as errors.
Fig. 6 shows a pattern of error detection that is strikingly similar to categorical
perception with a boundary in the vicinity of 5% voicing (Liberman, 1997). Note
that there is a heavy asymmetry in the perception of errors in /s/ and /z/ prod-
uctions. While an /s/ with almost any voicing was perceived as [z], few cases of
devoiced /z/ were perceived to be [s] or ½z  . Even when there was no voicing at all
in the /z/, it often sounded normal. Thus, while the production data suggest that
there were more anomalies in the /z/ tokens than the /s/ tokens, the perception of
this variation was heavily biased to detect /s/ errors but not /z/ errors. This
perceptual asymmetry explains the apparent reversal of the frequency effect that was
found by Stemberger (1991) and provides further evidence that speech error
transcription is an unreliable method for examining the speech production process.

3.2. Variation in frication amplitude and duration

Unusual variation in percent voicing reveals that there are many gradient errors in
the production of /s/ and /z/ in our corpus. It is less clear how the dimensions of
frication amplitude and duration might reveal gradient errors, since these dimensions
are not fully contrastive. Even though the distributions of frication amplitude and

Figure 6. Perceived error rate based on the authors’ percepts. Tokens of /s/
and /z/ are grouped across all participants for the same percent voicing groups
in Table I and Fig. 3. Each point shows the mean percent voicing for tokens
in the group (x-axis) and the percentage of tokens in the group that were
perceived as errors (y-axis).
154 S. A. Frisch and R. Wright

duration do not categorically differentiate /s/ from /z/, they do provide secondary
cues to the voicing distinction (Cole & Cooper, 1975; Shadle, 1985). We have shown
that tokens with 30% or more voicing were reliably perceived as [z]. Cases with less
voicing were much less reliably categorized. Closer examination of the data reveal
variation in frication amplitude and duration, in addition to voicing, affected the
perception of errors. In particular, for tokens with small amounts of voicing, high
frication amplitude and long duration were cues for /s/, while low frication
amplitude and short duration were cues for /z/.
Table II shows mean frication amplitude and duration for /s/ and /z/ depending
on whether they were perceived as errors or not. The top portion of the table
presents aggregate data for tokens with 0–5% voicing, the middle portion presents
aggregate data for tokens with 5–30% voicing, and the bottom portion presents
aggregate date for tokens with 30–100% voicing. While very few /s/ with 0–5%
voicing were perceived as errors, those that were appear to have lower frication
amplitude and shorter duration than those that were not (for amplitude t(347)=1.7,
p=0.09; for duration t(347)=1.6, p=0.06). For /z/ with 0–5% voicing, the pattern
was similar. Tokens with high frication amplitude and long duration were perceived
as [s], those with low frication amplitude and short duration were perceived as [z]
(for amplitude t(62)=4.0, po0.01; for duration t(62)=2.5, po0.01). For /s/ and /z/
with 5–30% voicing, frication amplitude and duration again appear to have an effect
on whether the tokens were perceived as errors or not, though the only statistically
significant difference is for frication amplitude for /s/ (t(22)=2.9, po0.01). Finally,
in cases of 30–100% voicing, there were no tokens that were not perceived as [z], so

TABLE II. Mean frication amplitude and duration for low, intermediate,
and high percent voicing groups as a function of intended segment and
perceived error for all participants

Group N Frication (dB) Duration (ms)

0–5% voicing
Perceived error 7 53.0 137.1
Non-error 342 56.3 173.2
Perceived error 25 54.5 163.2
Non-error 40 49.1 133.0
5–30% voicing
Perceived error 15 52.2 159.9
Non-error 10 57.1 148.8
Perceived error 2 53.0 131.2
Non-error 61 51.0 127.6
30–100% voicing
Perceived error 23 51.3 161.6
Non-error 0 F F
Perceived error 0 F F
Non-error 307 52.0 131.5
The phonetics of phonological speech errors 155

it appears that variation in frication amplitude and duration are not sufficiently
powerful cues to overwhelm the perceptual effect of periodicity.
If we compare the fully voiced /z/ tokens with the /s/ tokens with 0–5% voicing
that were perceived as errors, we see evidence that these voiceless /s/ tokens were
produced with frication amplitude and duration appropriate for /z/. These /s/ tokens
could reasonably be considered to be errors on the frication amplitude and duration
dimensions but not on the voicing dimension. In other words, these may be cases
where /s/ was produced as ½z . If this is the correct analysis, these tokens provide
additional evidence for gradient errors in production.

3.3. Other variation

3.3.1. Concatenation errors
There were 10 tokens, six for intended /s/ and four for intended /z/, where the
speakers appeared to correct their articulations without interrupting their produc-
tion. In all of these cases, the token begins as the unintended segment and is
corrected to the intended segment. In all cases, the resulting concatenation of onset
consonants is 300–500 ms long, and thus long enough to be equivalent to the
production of two separate segments. The resulting [zs] or [sz] onset cluster is, of
course, phonotactically illegal in English. Fig. 7 shows two examples of concat-
enation errors. The top panel shows [zs] in a production of sue by participant 6. The
bottom panel shows [sz] (or perhaps ½z z) for a production of zig by participant 3.

3.3.2. Other disfluencies

There are six other anomalous productions that are interesting to note. Three
of these productions appear to be linguistically explicable speech errors, but not
ones that were intended to be elicited by the experiment. All three cases involved

Figure 7. Example concatenation errors producing [zs] for sue (top panel) and
[sz] for zig (bottom panel).
156 S. A. Frisch and R. Wright

anticipation of the palatal articulation from the coda of the intended word zilch
onto the onset of the same word. In two cases, one by participant 6 and one by
participant 9, the result was [W]. Inspection of the spectrograms of these two
examples confirms that a voiced fricative was produced with frication noise in the
same region as the coda [pP]. Participant 6 interrupted his production mid-word and
immediately repeated the word correctly. In addition to being an unusual
anticipation, the phonotactic legality of the [W] onset is questionable in English.
The third palatal intrusion, also by participant 6, was combined with an error in
voicing. In this case, the palatal fricative was largely voiceless and of low amplitude.
The frication noise overlapped with the onset of the following vowel, resulting in a
short period of voiced frication. The perceptual result was [W]. In this case, the
participant once again interrupted his production mid-word, and then repeated the
word correctly after a 1 s delay.
The three other productions are a nonhomogenous set of disfluencies. These cases
are perceptually anomalous and could be considered speech production errors that
are gross gestural mis-coordinations. One case is a low amplitude ½z  like sound that
lasts for about 500 ms before becoming a relatively normal amplitude and duration
[s] onset. A second involves an intended /s/ where the alveolar constriction appears
to end about 100 ms before the onset of the following vowel, resulting in an
intrusive [h] percept. The third involves an intended /z/ that begins with about
100 ms of unfricated prevoicing followed by a 300 ms long fully voiced [z] onset.

3.3.3. ‘Geminate’ /s/

As noted above, one of the tongue twisters contained the sequence Zeus seem in
which a coda /s/ is followed by an onset /s/. There were 55 tokens of this sequence
across all participants. In 40 cases, the two fricatives were produced without a
discrete juncture. It is interesting to note that none of these long /s/ productions
show signs of being voiced and none were perceived to be errors. In addition, the 15
cases where there were two distinct /s/ productions also show very little variation in
voicing in the /s/ productions. Only one of the 15 distinct onset /s/ tokens has any
voicing, and that case involves just 3% voicing from a small amount of overlap with
the following vowel. Given that more than 15% of the other /s/ tokens contain
some amount of erroneous voicing, the doubled articulation of /s/ appears to have
somehow ‘‘protected’’ these segments from intrusion by /z/.
More general effects of segmental context on voicing of /s/ and /z/ (as well as /f/
and /v/) were found by Pirello, Blumstein & Kurowski (1997). They examined the
effect of segmental context on the voicing of fricatives in nonsense phrases like his
fav sips it. They found that voiceless fricatives following voiced segments were often
initially voiced, and voiced fricatives following voiceless segments were often initially

3.4. Summary
The concatenation errors and the many cases of intermediate amounts of voicing
found in all nine participants in our study supports the conclusions of Mowrey &
MacKay (1990). A significant number of tokens in our corpus are phonetically or
phonotactically abnormal in English. Thus, it appears that gradient errors do occur
The phonetics of phonological speech errors 157

frequently, and the process of phonetic accommodation does not regularize all
speech errors to be acoustically normal English sounds. We also found that the
gradient errors might or might not be detected as errors by careful listening.
On the other hand, there is also evidence for categorical effects in errors. Cases
where /s/ and /z/ are realized with a categorical change in voicing are more common
than we would expect if categorical changes in voicing were merely extreme exam-
ples of gradient voicing errors (i.e., the tails of the variation in the distribution of
voicing). In addition, there is clear evidence that many cases of 0% voicing for
/z/ also have duration and frication amplitude that is appropriate for /s/. It seems
likely that these are instances of categorical errors, where a normal articulation
appropriate to the wrong segment is produced. While there is some aerodynamic
interdependence between voicing, frication amplitude, and duration, their inter-
dependence is not so strong that a fricative of the length and amplitude of a normal
/s/ must necessarily be completely voiceless. In fact, our corpus contains many
productions of /z/ that have the length and frication amplitude of /s/ but are voiced
We conclude that a linguistic level of the segment or feature does have an
influence on the phonetic details of speech error production. This effect can be
captured in models of language processing that use graded activation and
competition between linguistic units to explain speech errors in phonological
encoding (e.g., Dell, 1986). Errors occur when competition results in the accidental
mis-selection of an incorrect word, segment, or feature. In order to account for
gradient speech errors, these models must be extended to include activation and
competition among phonetic articulatory plans. While it is certainly not a trivial
task to develop a working connectionist model of phonetic implementation, the
conceptual extension is a simple one. If there is activation and competition in
phonetic planning, there is the possibility for articulatory plans that mix gestures for
different articulators from different segments. Complete devoicing of /z/ with a ½z 
result could be explained by a combination of the glottal abduction of /s/ with
normal pulmonic effort associated with /z/. It is also conceivable that erroneous
articulatory plans that blend articulatory parameters for the same articulator could
be created. Partial voicing of /s/ could result from a combination of the articulatory
gestures for /s/ with the timing of oral and laryngeal gestures for /z/ in which the
fricative constriction is overlapped with the onset of voicing of the following vowel.
Activation and competition may also explain why ‘‘geminate’’ /s/ appeared to be
unaffected by /z/. Elements of articulatory planning, such as the laryngeal abduc-
tion, may be more strongly encoded when onset and coda /s/ productions come
together, as each segment would independently reinforce the activation of this
gesture. Correct /s/ voicing articulation will be more successful in competing with /z/
as the result of this additional activation.

4. Lexical effects

In the previous section, we argued that there are observable segmental effects in the
phonetics of speech errors. In our corpus, two of the tongue twisters were designed
so that the elicited errors would result in words and two of the tongue twisters were
designed so that the elicited errors would result in nonwords. We, therefore, also
158 S. A. Frisch and R. Wright

have the opportunity to examine whether the productions were influenced by the
presence of a word error outcome. In the speech error literature, it has been
reported that word error outcomes are more likely than nonword outcomes in both
naturally occurring error corpora and in experiments that elicit speech errors
(Motley & Baars, 1975; Dell & Reich, 1980; Stemberger, 1983).
Table III shows the tokens divided into groups by percent voicing, as in Table I,
but with the tokens further divided by the lexical status of the error outcome of the
tongue twister. In the case of /s/, it appears that all types of error in voicing were
more common in the case of lexical outcomes, and the two distributions deviate
significantly from a common distribution (w2(6)=19.7, po0.01). This provides clear
evidence that the lexicality of the outcome affects both gradient and categorical
aspects of speech errors.
In the case of /z/, the pattern is less clear, though the distributions are
significantly different (w2(6)=13.9, po0.05). It appears that the number of tokens
with 0–30% voicing is somewhat higher in the lexical case than in the nonlexical
case, and the pattern is reversed for voicing from 30% up to 100%. As mentioned
in Section 3, it may be the case that the distribution of percent voicing for /z/
reflects a mixture of normal devoicing and error-induced devoicing. The number of
cases with highly reduced voicing in the lexical case may reflect a greater amount of
error-induced devoicing on those tokens. The fact that there are fewer lexical cases
than nonlexical cases with smaller amounts of devoicing is more difficult to explain.
One possibility is that normally devoiced productions of /z/ are somehow more
susceptible to additional error-induced devoicing than cases where /z/ is 100%
voiced. If error-induced devoicing is stronger in the lexical case, then the number of
tokens with only a small amount of devoicing would be differentially reduced.

TABLE III. Distribution of intended /s/ and /z/

tokens by percent voicing and lexical status of
potential error outcome for all participants

Percent voicing Lexical Nonlexical

0% 162 162
0–5% 20 5
5–10% 7 6
10–30% 11 1
30–60% 3 1
60–100% 5 1
100% 9 4
Total 217 180

0% 34 22
0–5% 4 4
5–10% 15 9
10–30% 23 16
30–60% 7 16
60–100% 10 23
100% 129 123
Total 222 213
The phonetics of phonological speech errors 159

Given the clear lexical effects on /s/ production and some indication of lexical
effects on /z/ production, we conclude that productions in which there is a com-
peting lexical outcome produce greater numbers of errors. The effect of a competing
lexical item on speech errors is found in an increase in both gradient errors and
categorical errors. Together with the evidence from Section 3 that categorical errors
in voicing are more common than expected based on the pattern of gradient errors,
we conclude that the segment and word are both units of organization in the speech
production process that affect the phonetic details of articulation.
The lexical effect we found in speech production further supports a connectionist
approach to phonological encoding in speech production. If the articulatory plans
for segments compete with one another during encoding, then competing segments
that are reinforced by corresponding word nodes will be enhanced. For competing
word outcomes, more errors will occur. When there is no word competitor, the
correct segment will encounter less competition, and fewer errors will occur.
Assuming that errors can involve both blending of articulatory plans and wholesale
substitution, an increase in both gradient and categorical errors would be predicted
in cases where there is a lexical competitor.

5. Discussion and conclusion

This study presented an acoustic and perceptual analysis of onset /s/ and /z/ speech
errors by nine talkers. In support of the claims of Mowrey & MacKay (1990), we
found that gradient, noncontrastive errors can occur, and that such errors are
actually common. In addition, we found that categorical errors also occur at rates
that are higher than would be expected if the only source of errors was from
noncontrastive variation that happened to extend into another phonetic category.
Finally, we demonstrated a lexical effect on both gradient and categorical errors.
These patterns provide evidence for a set of higher level units that organize phonetic
gestures at the level of the segment and word, agreeing with some of the obser-
vations of traditional speech error analyses based on transcribed data.
While some of the theoretical conclusions of transcription-based speech error
analyses are supported, others must be rejected. Assertions that speech errors result
in grammatically acceptable utterances are not supported (e.g., Fromkin, 1971). The
detailed acoustics of some of the productions in our corpus, specifically those
resulting in intermediate voicing errors, are distinctly phonetically anomalous. For
example, errors where 10% of an intended /s/ is voiced by overlapping with the
following vowel are clearly anomalous, and by no means similar to cases of
normally devoiced /z/. In addition, the concatenation errors are instances of
phonotactically ungrammatical onset sequences. While self-monitoring and editing
certainly takes place in speech production (Levelt, 1989), it is not the case that the
majority of speech errors in our corpus are best analyzed as mis-selection of
phonological segments that have been phonetically encoded after the error occurred.
These findings support the claims of a growing number of researchers that
transcription is inadequate for complete error coding, as transcription makes
incorrect assumptions about the wholly categorical and abstract nature of the data
(Laver, 1980; Boucher, 1994; Ferber 1995). Transcription techniques are susceptible
to perceptual bias, and so can be expected to be effective only when bias is
160 S. A. Frisch and R. Wright

minimized. Thus, data from experimentally elicited errors that are recorded and can
be examined repeatedly are much more reliable than naturally occurring errors that
are transcribed opportunistically while the ‘‘experimenter’’ is engaged in conversa-
tion or otherwise not completely focused on the transcription task. But even careful
listening techniques will still be influenced by listener bias, and gradient errors are
likely to be missed. In our own perception of the tokens in our corpus, there was a
heavy perceptual bias to detect errors in /s/ production and not errors in /z/
production. The acoustically determined error pattern suggests that there were in
fact more /z/ errors than /s/ errors.
The categorical tendencies in the acoustics of the productions in our corpus
provide evidence that gestures tend to be organized into permissible segments. This
tendency interacted with the lexical status of the outcome, suggesting that gestures
tend to be organized into permissible words. Extending the tendency to produce
phonetically normal segments and words to the more general case can explain the
traditional claim that errors result in grammatical sequences of segments. Errors
tend to result in grammatical sequences. Stemberger’s (1983) very carefully collected
corpus of naturally occurring errors supports the claim that grammaticality is a
tendency, but not an absolute. He found some perceptually detectable but
phonotactically anomalous productions in naturally occurring dialogue. The
agreement between our analysis and Stemberger’s corpus also suggests that our
conclusions are not specific to tongue twisters, and that they apply to speech
production in general.
Overall, our findings are incompatible with models of speech production that
involve selecting, organizing, and ordering abstract discrete representational units
into a frame that is then implemented by the phonetic module (e.g., the scan-copier
model of Shattuck-Hufnagel, 1979). The interaction between the articulatory details
of speech production and organization at the lexical and segmental levels suggests
that there is a hierarchically organized and interactive production mechanism that
manipulates phonetic units directly. Appropriate models that use discrete linguistic
levels and spreading activation (Dell, 1986) or less explicitly organized connectionist
architecture (Dell, Juliano & Govindjee, 1993) have previously been proposed as
models of phonological encoding that sometimes generate speech errors at the level
of words and segments. These models are compatible with our findings if they are
extended to include competition at the level the articulatory plans of segments. We
conclude from this that the phonetics of phonological speech errors is an important
topic for future research in speech production.

Parts of this research were presented at the 133rd and 134th meetings of the Acoustical Society of
America, the 72nd annual meeting of the Linguistic Society of America, and the University of Illinois at
Urbana/Champaign. We would like to thank those audiences for their questions and suggestions. We
would also like to thank Dani Byrd, Terry Nearey, and two anonymous reviewers for helpful comments
on an earlier draft of this paper
This work supported in part by NIH Training Grant DC 00012 to Indiana University and by the
Language Learning Research Assistant Professorship at the University of Michigan.


Baars, B., Motley, M. & MacKay, D. (1975) Output editing for lexical status from artificially elicited slips
of the tongue, Journal of Verbal Learning and Verbal Behavior, 14, 382–391.
The phonetics of phonological speech errors 161

Baum, S. & Blumstein, S. (1987) Preliminary observations on the use of duration as a cue to syllable-
initial fricative consonant voicing in English, Journal of the Acoustical Society of America, 82,
Boucher, V. (1994) Alphabet-related biases in psycholinguistic enquiries: considerations for direct theories
of speech production and perception, Journal of Phonetics, 22, 1–18.
Byrd, D. (1996) A phase window framework for articulatory timing, Phonology, 13, 139–169.
Cole, R. (1973) Listening for mispronunciations: a measure of what we hear during speech, Perception &
Psychophysics, 13, 153–156.
Cole, R. & Cooper, W. (1975) The perception of voicing in English affricates and fricatives, Journal of the
Acoustical Society of America, 58, 1280–1287.
Dell, G. (1986) A spreading activation theory of retrieval in sentence production, Psychological Review,
93, 283–321.
Dell, G., Juliano, C. & Govindjee, A. (1993) Structure and content in language production: a theory of
frame constraints in phonological speech errors, Cognitive Science, 17, 149–195.
Dell, G. & Reich, P. (1980) Toward a unified theory of slips of the tongue. In Errors in linguistic
performance: slips of the tongue, ear, pen, and hand (V. Fromkin, editor), New York: Academic Press.
pp. 273–286.
Ferber, R. (1995) The reliability and validity of slip-of-the-tongue corpora: a methodological note,
Linguistics, 33, 1169–1190.
Frisch, S. (1996) Similarity and frequency in phonology. Unpublished PhD dissertation, Northwestern
University, Evanston, IL.
Fromkin, V. (1971) The non-anomalous nature of anomalous utterances, Language, 47(1), 27–52.
Garnham, A., Shillcock, R. C., Brown, G. D. A., Mill, A. I. D. & Cutler, A. (1982) Slips of the tongue in
the London-Lund corpus of spontaneous conversation, Linguistics, 19, 805–817.
Haggard, M. (1978) The devoicing of voiced fricatives, Journal of Phonetics, 6, 95–102.
Klatt, D. (1976) Linguistic uses of segmental duration in English: acoustic and perceptual evidence,
Journal of the Acoustical Society of America, 59, 1208–1221.
Laver, J. (1980) Slips of the tongue as neuromuscular evidence for a model of speech production.
In Temporal variables in speech (H. Dechert & M. Raupach, editors), The Hague: Mouton.
pp. 21–26.
Levelt, W. (1989) Speaking: from intention to articulation. Cambridge, MA: MIT Press.
Liberman, A. (1997) Speech: a special code. Cambridge, MA: MIT Press.
Marslen-Wilson, W. & Welsh, A. (1978) Processing interactions and lexical access during word recognition
in continuous speech, Cognitive Psychology, 10, 29–63.
Mowrey, R. & MacKay, I. (1990) Phonological primitives: electro-myographic speech error evidence,
Journal of the Acoustical Society of America, 88(3), 1299–1312.
Motley, M. & Baars, B. (1975) Encoding sensitivities to phonological markedness and transition
probability: evidence from spoonerisms, Human Communication Research, 2, 351–361.
Ohala, J. J. (1983) The origin of sound patterns in vocal tract constraints. In The production of speech
(P. F. MacNeilage editor), New York: Springer-Verlag. pp. 189–216.
Pirello, K., Blumstein, S. E. & Kurowski, K. (1997) The characteristics of voicing in syllable-initial
fricatives in American English, Journal of the Acoustical Society of America, 101, 3754–3765.
Pickett, J. (1980) The sounds of speech communication. Baltimore, MD: University Park Press.
Pouplier, M., Chen, L., Goldstein, L. & Byrd, D. (1999) Kinematic evidence for the existence of gradient
speech errors, Journal of the Acoustical Society of America, 106, 2242.
Saltzman, E. L. & Munhall, K. G. (1989) A dynamical approach to gestural patterning in speech
production, Ecological Psychology, 1, 333–382.
Samuel, A. G. (1981) Phonemic restoration: insights from a new methodology, Journal of Experimental
Psychology: General, 110, 474–494.
Shadle, C. (1985) The acoustics of fricative consonants. Cambridge, MA: MIT Press.
Shattuck, S. (1975) Speech errors and sentence production. Unpublished PhD dissertation, Massachusetts
Institute of Technology.
Shattuck-Hufnagel, S. (1979) Speech errors as evidence for a serial-ordering mechanism in sentence
production. In Sentence processing: Psycholinguistic studies presented to Merrill Garrett (W. E. Cooper
& E. C. T. Walker, editors), Hillsdale, NJ: Erlbaum. pp. 295–342.
Shattuck-Hufnagel, S. (1983) Sublexical units and suprasegmental structure in speech production planning.
In The production of speech (P. F. MacNeilage, editor) New York: Springer-Verlag. pp. 109–136.
Shattuck-Hufnagel, S. (1992) The role of word structure in segmental serial ordering, Cognition, 42,
Smith, C. L. (1997) The devoicing of /z/ in American English: effects of local and prosodic context,
Journal of Phonetics, 25, 471–500.
Stemberger, J. P. (1983) Speech errors and theoretical phonology: a review. Bloomington, IN: Indiana
University Linguistics Club.
162 S. A. Frisch and R. Wright

Stemberger, J. P. (1991) Apparent anti-frequency effects in language production: the addition bias and
phonological underspecification, Journal of Memory and Language, 20, 161–185.
Strevens, P. (1960) Spectra of fricative noise in human speech, Language and Speech, 3, 32–49.
Warren, R. M. (1970) Perceptual restoration of missing speech sounds, Science, 176, 392–393.
Wells, R. (1951) Predicting slips of the tongue, Yale Scientific Magazine, December, 9–12.
Wickelgren, W. (1965) Distinctive features and errors in short-term memory for English vowels, Journal of
the Acoustical Society of America, 38, 583–588.
Wright, R., Frisch, S. & Pisoni, D. B. (1999) Speech perception. In Encyclopedia of electrical and
electronics engineering (J. Webster, editor), New York: John Wiley & Sons. pp. 175–195.